Every trading bot marketplace showcases equity curves that climb smoothly upward, yet most buyers discover a different reality once real money is at stake. The gap is rarely fraud in the narrow sense; it is backtesting methodology. A backtest that uses coarse candles, ignores fees and slippage, or tunes parameters to one lucky historical window will make almost any strategy look profitable. Learning to spot those flaws protects you twice: you avoid buying bad bots, and you avoid trusting your own tests when you build one.

The most common flaw is testing on hourly or daily candles when the bot trades on minute-level signals. If a grid level sits inside a single hourly candle's range, a naive backtest may book a full buy-and-sell round trip inside that bar, even though in live trading the order of touches inside the candle determines whether both sides filled at all. The honest fix is to test on one-minute candles or actual trade data, and if only coarse data exists, to adopt a conservative rule: at most one fill per level per bar, and never a buy and its matching sell in the same bar.

Look-ahead bias is the quieter cousin of this error. It happens when a backtest accidentally uses information that was not available at decision time, such as a daily close to trigger an intraday trade, or a revised economic figure instead of the original release. Reputable testing frameworks prevent this by construction, but spreadsheet backtests rarely do. Before trusting any result, ask exactly what data each decision used and whether that data existed at that timestamp in real life.

A backtest that omits fees is not a backtest at all, it is a fantasy. High-frequency and grid strategies can generate hundreds of trades whose individual profits are smaller than the commission plus spread, turning a glowing gross result into a net loss. Always model maker and taker fees for your actual exchange tier, add a slippage assumption of at least a few basis points per trade, and include funding payments for futures strategies. Then compare the bot against simply buying and holding the same asset, because a bot that underperforms holding while adding complexity has failed its basic purpose.

Overfitting is the third great flatterer. If you test 50 combinations of parameters and pick the best, part of that winner's performance is luck: the parameters were chosen because they fit that particular history. Warning signs include results that collapse when you shift a parameter slightly, or a strategy that works in one three-month window and fails in the next. The defence is out-of-sample testing: optimise on one period, then verify on a later period the parameters never saw, and require the strategy to survive across sideways, trending up, and trending down regimes.

Judge a backtest on equity including open positions, marked to market every bar, not on closed-trade win rate. Grid bots famously show 90 percent win rates on closed trades while carrying large unrealised losses in open inventory, so win rate alone is nearly meaningless. The metrics that matter are net profit after all costs, maximum drawdown including unrealised losses, profit divided by worst drawdown, and the number of genuinely independent market regimes tested. Three hundred round trips inside one sideways month are closer to one observation of a sideways market than to 300 independent tests.

Finally, demand transparency from any bot seller: the exact test period, the fee and slippage assumptions, the out-of-sample results, and the drawdown figures. A seller who shows only a rising equity curve and an annual percentage yield is selling marketing, not evidence. Run the same checklist on your own strategies before risking capital, and you will discard most bots quickly, which is precisely the point: in automated trading, saying no to 95 percent of strategies is the skill that preserves capital for the few that deserve it.

Honest backtesting is slower and less exciting than a perfect equity curve, but it is the only kind that predicts live results. Test on fine data, charge every fee, validate out of sample, and judge on net drawdown-adjusted returns. A bot that survives that gauntlet is worth a small live trial; one that needs shortcuts to look good is not.